Methods for Deriving Pesticide Aquatic Life Criteria - Reviews of Environmental Contamination and Toxicology - page 80

(ANZECC and ARMCANZ 2000) did not adopt the USEPA SSD approach (1985),
because its data requirements are too stringent, there is no biological basis for
selecting a triangular distribution, not all of the data are used, and it assumes that a
threshold toxicity value exists. In defense of the USEPA approach (1985), Erickson
and Stephan (1988) argue that, because the entire data set is used in setting percen-
tile ranks and cumulative probabilities, calculation of the FAV using the four data
points nearest the 5th percentile does not constitute “not using all the data.” Their
interpretation is that, those four data points are used as a means of giving more
weight to toxicity values nearest 5th percentile. This weighting, however, leads to
other problems, which are discussed below.
To sidestep the issue of choosing an appropriate distribution, several researchers
believe it best to make no assumptions about distribution shapes, and to use nonpara-
metric methods to estimate community or ecosystem effects, where they are based
on single-species toxicity tests (Jagoe and Newman 1997; Van Der Hoeven 2001;
Grist et al. 2002). Grist et al. (2002) found that HC 5 values, determined by parametric
versus nonparametric methods, were significantly different from each other.
Wheeler et al. (2002) suggest that, to get the best HC 5 (or, generally, HC p ) estimate,
data should be analyzed by four different SSD methods (two parametric and two
nonparametric), selecting the one that gives the best fit. While bootstrapping, techniques
offer a solution to the distribution problem, they are very data intensive, and will
not work for many small data sets available for criteria derivation.
Arguments for one or the other distribution, or for making no distributional
assumptions, are based on which distributions are easier to work with, or which
ones better quell the criticism that SSDs are not valid, because data usually do not
fit the assumed distribution. Ultimately, all methods currently in use appear to
derive protective criteria. In the Netherlands, the log-normal distribution was
selected over a log-log distribution (Aldenberg and Slob 1993), because the distri-
butions are not so different, results obtained are not different, and the normal
distribution provides powerful statistical tools (RIVM 2001). Similarly, the OECD
(1995) concludes that the log-normal, log-logistic, and triangular distribution methods
give very similar results.
The ANZECC and ARMCANZ guidelines (2000) take the data-fitting idea a
step further in a modification of the Dutch approach. In the Australia/New Zealand
methodology data are fitted to one of a family of Burr distributions (Burr 1942;
Shao 2000), and then the HC 5 extrapolation is performed using the best fit. This approach
allows for derivation of high and moderate reliability TVs from data that would
have precluded using log-normal or log-logistic distributions. Noting that the Dutch
(MHSPE 1994; RIVM 2001) and Danish (Samsoe-Petersen and Pedersen 1995)
SSD methodologies give very similar results, and differ only in the selection of
either a log-logistic (Dutch; Aldenberg and Slob 1993) or log-normal (Danish;
Wagner and Løkke 1991) distribution, the Australia/New Zealand guidelines chose
to start with the Dutch approach, because it had been more extensively evaluated
and was easier to use. Advantages to the Dutch approach include that it uses the full
range of available data, and a water manager can choose a level of protection and
a level of uncertainty associated with a guideline value.